OpenCodex: a local multi-terminal workspace for AI coding

The AI coding desktop I wanted was straightforward: open a folder, split the screen into terminals, and run Claude Code or Codex. Rearranging the layout should keep work running; closing a terminal should clean up its child processes. OpenCodex was built around that workflow.

The original article described a demo using Tauri and the system WebView, with a package of about 4.7 MB. That is a historical build size, not a fixed size for every release or platform. See the OpenCodex repository for code, prerequisites and releases.

Folders and terminals organize the work

The project folder is the starting point, with horizontal or vertical terminal splits. One terminal can develop while another builds. Splitting changes layout and should not terminate a running task.

Multiple terminals still need ownership of files. Layout does not resolve two AIs editing the same file. Shared routes, dependencies and version files require coordination, and Git differences still need review.

Three practical details

• Preserve processes: rearranging a split must not kill a running Claude terminal.
• Clean up process trees: closing a terminal should handle its children rather than leave orphaned Node processes.
• Pass real paths: dragging files, images or folders inserts their paths into the command line.

The historical version also recorded layout restoration, adjustable text size and WebGL rendering. Performance and compatibility still need checking on your machine.

A local client and a model service are separate

OpenCodex has no embedded API key and does not require my own server. Users configure cloud or local endpoints. The workspace can run locally, but cloud requests go to the configured service and may incur charges.

Check endpoint, model, authentication and data destination before use. A local workspace does not automatically mean inference is offline or free. Do not commit secrets to project files.

A sensible first session

1. Prepare the required WebView and command-line tools.
2. Open a test project and check one terminal starts and exits.
3. Add splits and check running work survives layout changes.
4. Configure your endpoint and call it with non-sensitive material.
5. Use a real project with an editing scope for each terminal.

OpenCodex and Redline are separate projects

OpenCodex is MIT-licensed. The original article also introduced Redline, an Apache-2.0 annotation layer whose historical record described 11 formats. A person marks regions, arrows and comments; AI uses them to edit source files. Redline itself does not write back to the source. It is not another name for OpenCodex.

Continue with U-King's development practice or my Chinese Codex CLI tutorial.

By Hequbing (He Fangsheng / Dosen). English edition prepared on 2 October 2026, based on my WeChat article (Chinese). Original illustrations and publication records are available there.

Contact: WeChat hecare888; hefangsheng@gmail.com. Services and pricing.